lets try functiongemma instead

train/chat_templates/gemma3_withtools.j2

@@ -2,7 +2,7 @@
 {%- if not tools or tools | length == 0 %}No tools were provided. If the user requests you interact with a device, tell them you are unable to do so.{% else %}
 Tools:
 {% for tool in tools %}
-- {{ tool['name'] }}({{ tool['parameters']['properties'].keys() | join(', ') }}): {{ tool['description'] }}
+- {{ tool['function']['name'] }}({{ tool['function']['parameters']['properties'].keys() | join(', ') }}): {{ tool['function']['description'] }}
 {% endfor -%}
 {%- endif -%}
 {%- for message in messages -%}
@@ -13,8 +13,7 @@ Tools:
     {%- else -%}
         {%- set role = message['role'] -%}
     {%- endif -%}
-    {{ '<start_of_turn>' + role + '
-' }}
+    {{ '<start_of_turn>' + role + '\n' }}
     {%- if role == "tool" -%}
         {{ '<tool_result>' }}
     {%- endif -%}
@@ -41,13 +40,7 @@ Tools:
     {%- endif -%}
     {%- if message['tool_calls'] is defined and message['tool_calls'] | length > 0 %}
         {%- for tool_call in message["tool_calls"] -%}
-            {{ '\n<tool_call>{"name": "' + tool_call['name'] + '", "arguments": ' + ('"' + tool_call['arguments'] + '"' if tool_call['arguments'] is string else tool_call['arguments'] | tojson) + '"}</tool_call>' }}
+            {{ '\n<tool_call>{"name": "' + tool_call['function']['name'] + '", "arguments": ' + ('"' + tool_call['function']['arguments'] + '"' if tool_call['function']['arguments'] is string else tool_call['function']['arguments'] | tojson) + '"}</tool_call>' }}
         {%- endfor %}
     {%- endif -%}
-    {{ '<end_of_turn>
-' }}
-{%- endfor -%}
-{%- if add_generation_prompt -%}
-    {{'<start_of_turn>model
-'}}
-{%- endif -%}
+    {{ '<end_of_turn>\n' }}

train/functiongemma-270m.yml

@@ -0,0 +1,43 @@
+base_model: google/functiongemma-270m-it
+model_type: Gemma3ForCausalLM
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+datasets:
+  - path: /workspace/data/datasets/sample.jsonl
+    ds_type: json
+    type: chat_template
+    roles_to_train:
+      - assistant
+
+val_set_size: 0.0
+output_dir: /workspace/data/training-runs/Home-Gemma3-270m
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+
+use_tensorboard: true
+
+# batch size = 16
+gradient_accumulation_steps: 16
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0

train/gemma3-270m.yml

@@ -10,7 +10,7 @@ chat_template_jinja: |
   {%- if not tools or tools | length == 0 %}No tools were provided. If the user requests you interact with a device, tell them you are unable to do so.{% else %}
   Tools:
   {% for tool in tools %}
-  - {{ tool['name'] }}({{ tool['parameters']['properties'].keys() | join(', ') }}): {{ tool['description'] }}
+  - {{ tool['function']['name'] }}({{ tool['function']['parameters']['properties'].keys() | join(', ') }}): {{ tool['function']['description'] }}
   {% endfor -%}
   {%- endif -%}
   {%- for message in messages -%}
@@ -21,8 +21,7 @@ chat_template_jinja: |
       {%- else -%}
           {%- set role = message['role'] -%}
       {%- endif -%}
-      {{ '<start_of_turn>' + role + '
-  ' }}
+      {{ '<start_of_turn>' + role + '\n' }}
       {%- if role == "tool" -%}
           {{ '<tool_result>' }}
       {%- endif -%}
@@ -49,15 +48,13 @@ chat_template_jinja: |
       {%- endif -%}
       {%- if message['tool_calls'] is defined and message['tool_calls'] | length > 0 %}
           {%- for tool_call in message["tool_calls"] -%}
-              {{ '\n<tool_call>{"name": "' + tool_call['name'] + '", "arguments": ' + ('"' + tool_call['arguments'] + '"' if tool_call['arguments'] is string else tool_call['arguments'] | tojson) + '"}</tool_call>' }}
+              {{ '\n<tool_call>{"name": "' + tool_call['function']['name'] + '", "arguments": ' + ('"' + tool_call['function']['arguments'] + '"' if tool_call['function']['arguments'] is string else tool_call['function']['arguments'] | tojson) + '"}</tool_call>' }}
           {%- endfor %}
       {%- endif -%}
-      {{ '<end_of_turn>
-  ' }}
+      {{ '<end_of_turn>\n' }}
   {%- endfor -%}
   {%- if add_generation_prompt -%}
-      {{'<start_of_turn>model
-  '}}
+      {{'<start_of_turn>model\n'}}
   {%- endif -%}
 special_tokens:
   eot_tokens: